A trajectory prediction method based on continual learning with dynamic adjustment of feature weights

Due to the heterogeneous interactions among multiple traffic participants in the intelligent driving scenario and its dynamic coupling with the road topology structure, high-precision trajectory prediction remains a key challenge in intelligent driving systems.Existing rule-based and deep learning approaches for trajectory prediction continue to grapple with several key limitations: the inability to assess the complexity of hierarchical driving scenarios in real time, the reliance on static feature fusion strategies that fail to adapt to environmental dynamics, and the absence of continual learning mechanisms tailored for evolving traffic conditions. To address these challenges, this paper proposes a trajectory prediction method based on continual learning with dynamic feature weight adjustment. Firstly, a hierarchical environmental complexity quantification model is proposed, which dynamically evaluates driving environment complexity features by assessing local and global complexity. Secondly, a GRU-driven feature weight adjustment coordination module is constructed, which automatically adjusts the local-global feature weights in the trajectory prediction task based on real-time environmental complexity. Compared with baseline methods, this achieves a reduction of 4.4% in minFDE and 3.1% in minADE. Finally, the proposed continual learning module, integrating weight consolidation, dynamic priority experience replay, and online adaptive inference, is evaluated through extensive experiments on the Argoverse 1.1 dataset. The results demonstrate its superior performance, particularly in high-complexity scenarios. Compared to baseline methods, the proposed method achieves the lowest minFDE (1.809 m) and minADE (1.036 m), with a miss rate of only 0.253. The proposed method offers a paradigm for intelligent driving systems to handle trajectory prediction under dynamically complex environments, effectively enhancing the long-term learning and stable prediction capabilities of trajectory prediction models.